From: AAAI Technical Report SS-94-04. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Active Rescheduling for Goal Maintenancein DynamicManufacturing- Systems ArneF. Claassen, RubyD. Lathon, Daniel M. Rochowiak, Leslie D. Interrante Intelligent SystemsLaboratory University of Alabama in Huntsville intert@ebs330.eb.uah.edu Abstract organizationbeginsto crumbleas the tasks at hand begin to pull frommultiplebureaucratic agencies, The manufacturingplant is viewedas a group of and the abstract organizational plan meets interacting subsystems.Theresearch describedin recalcitrant realities. The constantly changing this paperis aimedat characterizingthe subsystem workplacecreates demandson the most immediate interactions such that an intelligent systemcan bureauto changeits plans, if there is to be hopeof makeappropriatedecisionsin the face of shopfloor meetingits goal. Further, the demands of multiple contingencies.Thegoal of such decision makingis bureaucraticagenciesintersect in the workplace, to insure the meetingof global productiongoals by and the multiple demandswill often, but not makingreal-time tradeoffs at the level of local always,require that onebureaucraticagencyinteract subsystem goals. Manufacturing subsystems withanother,if either agencyis to achieveits goal include receiving, shipping, production lines, (BondandGasser,1988;Durfeeet al., 1989). maintenance,and material handling. It is assumed that subsysteminteractions can be captured as B. The Challenge scheduleinteractions. Activereschedulingis the use of manufacturing parameters as "cues" for The constant tension between the a priori triggering subsystemrescheduling. A distributed bureaucraticstructure and the a posteriori needto agent system is described, in which an agent adapt to newand unforeseencircumstancesleads to represents a subsystemscheduler. Twoagendasare a collision of organizational structure and work employed:onefor the posting of proposedschedule place dynamism that is especially apparent in the revisions anda secondagendafor the processingof manufacturingindustry. Thechallenge is to build contingencies which must be resolved before an architecture that represents both the proposedschedulerevisionscan be accepted. organization’sstatic structure andthe workplace’s dynamism.Theorganization’s static structure is I. Goal Maintenance in a Dynamic encapsulatedin the agencygivento its bureaucratic Environment units. Its units havethe powerto respondto certain changesthroughprescribed rules used to maintain A. The Issue goal states. However, work place events may requirethat a unit negotiatewith other units in the The modern world has created numerous effort to maintainits goal state. In suchcases, the bureaucraciesto carry out specific local goals that rules of the agentare not sufficient withinits scope are believedto contribute to the satisfaction of a of agencyto handle the situation (Hewitt, 1986; global goal. Atypical organizationis dividedinto Star, 1989). bureausthat are eachchargedwith a particular task. Wepropose an architecture in whichmultiple Eachbureau has a specific goal and its actions agents participate in a hierarchy, havea degreeof focus on the satisfaction of that goal. Fromthe autonomyin taking local actions to satisfy local organization’spoint of view, the satisfaction of goals, respond to changing conditions, can eachbureau’sgoal will producethe satisfaction of negotiatewith other agentsin attemptingto satisfy the global goal. This authoritarianstructure can be their local goals, andare subject to the judgments repeated as manytimes as necessaryto create an of the agentthat is focusedon the satisfaction of enterprise where the satisfaction of the most global goals. For our purposes,an agent represents specific goals will producethe satisfaction of the the agencyof a unit in an organization.It receives most global goal. Eachbureausets morespecific informationandcan vary its operationby reasoning demandson the bureaus that fall under it by about the current state and its goal. Further, the creatingincreasinglyspecific goals. agent can enter into negotiationwith other agents, Opposed to this hierarchicalbureaucraticstrategy if it is unableto maintainits goal state through is the reality of the workplace. Theworkplace is local changes commensuratewith its agency. In constantly changing and placing demandson the this sense, agents represent organizationalagency immediate bureaucracy which is attempting to and not the psychologicalagencyof the human. satisfy its goal. In the workplace,the bureaucratic 33 Agentsembodyknowledgeand have resources for communication and negotiation. The latter characteristic separates themfrom a traditional knowledge-basedsystem. Each agent has its own information and knowledge which guides its actions. Each agent can inform other agents or request resources. For any particular agent, the sensors, databases, and actions of other agents constitute the external world.Theagent’sinternal world is constituted by the knowledge and understanding that it has of its particular task. The spectrumof organizational modelsof agents in a systemspans two poles, fromstrictly hierarchical (agent results reported to a higher level where authority exists to makedecisions)to planar (agent results reportedto all other agents; all haveequal authority). Ourintelligent systemattemptsto avoid the excesses of these poles. Unlike the focus of attention in a hierarchical system, agents do take accountof the rest of the world.Unlikethe focusof attention in planar systems,agents are constrained in the actions that they maytake. In brief, intelligent focusof attention requiresthat agentsbe awareof their ownresources,goals, andconstraints in the effort to solve a problemthat potentially affects the entire collectionof agents(Interrante et al, 1993a). C. Agents and Schedules in Manufacture Achievement of the global goals of manufacturing requires an understandingof interacting schedules and tradeoffs. Manufacturing research has produced manyanalytic techniquesfor generatingindividual schedules. Typically, however,interactions with other schedules(with the exceptionof the master production schedule) are handled by making assumptionsthat create boundaries.Thescheduleof interest is thus separated from the schedulesof other agents andthe dynamism of the shopfloor. A scheduler agent views the plant as a closed, idealized world,in whichcause-and-effectanalysis across organizational units is ignored. Bounding assumptions keep the plans and schedules in relative isolation, and embodyassumptionsabout the abilities of neighboring subsystems. The application of rigorous analytical techniques is normallytightly constrainedby these assumptions and problemsoccur whenshopfloor events violate them. Theconflict of agents is a result of the closedworld assumption that any knowledge not explicitly representedin the agent nor inferable from such knowledgeis irrelevant to the agent’s activity. The closed world assumption can be mitigatedby providinghigher-orderrepresentations and reasoningto moderateconflicts amonglowerlevel agents. The closed world is openedto some 34 degree by allowing more heuristic styles of reasoni-~gto moderateconflicts andviolations of the boundaries.Suchheuristic techniquesare also subject to the closed world assumption,but when combined with the analytical techniquesin a multiagentsystema "larger" closedworldis created. II. Negotiating Schedules: An Architecture A. Agents 1. Description The architecture described in this section in designed to capture manufacturing subsystem interactions such that appropriatedecisionscan be madein the face of shop floor contingencies. Subsysteminteractions are viewedas schedule interactions for our purposes. Toaid intelligent rescheduling the rescheduling actions should be performedin reaction to systemparameters(Smith et al. 1991)and the schedulingmodelshouldadhere to the physical modelof the plant (McKay,1993). Our system includes intelligent, autonomous agentsthat negotiateschedulechangesaffecting the entire system. This design is similar to one discussedby Burkeand Prosser (1991). In the case that a proposedschedulechangehas no real "cost" in termsof inhibiting the goalsof other agents, the change is implemented with no need for negotiation. Otherwise,negotiation amongagents occurs to determine whether a proposedschedule changeis appropriatein light of its contributionto the global masterproductionscheduleand its cost in terms of the effect on related subsystem capabilities. Unlike bidding schemes, which attemptto quantifysuchcontributionsand/orcosts, an overseeing agent employsheuristic schedule interaction knowledgeto determine whether a proposedschedule changeshould be implemented. Informally, such negotiation can be viewedas a combinationof "taking the easy road" and "doing the right thing". Eachagent is autonomous in that it attempts to protects its beliefs about howits schedulesshould be organizedand executed. Each agent represents a manufacturing system(receiving, material handling, productionlines, maintenance, andshipping) and has its ownset of schedulesand schedulingstrategies for performingthe necessary rescheduling. In this way, an agent is a selfcontained scheduling unit with the capability of communicating with the other agents. As in Zlotkin and Rosenschein (1991)the negotiation can result in either one of the agents involvedin the conflict yielding to the other agent, or a compromiseof the agent’s goals to promote a global goal. Unlike Zlotkin and Rosenscheinthis system utilizes a Superagent to oversee this stochastic simulation is used as a knowledge process. acquisition tool for the schedulingagents, similar TheSuperagentis similar to the other agents in to the approachof Nakasukaand Yoshida(1992). that it contains a schedule and a scheduling Theschedulingsystemis dividedinto three parts as strategy, but it is uniquein that its scheduleis the illustrated below: 1. a Superagent monitoring masterproductionschedulewhichis preeminentand interactions and analyzing shop floor data, 2. a embodies the global system goals. It has committee of subagents that represent the knowledge of agent interactions and the ability to production system and 3. a shop-floor database adjudicateconflicts among agents. Adiscrete-event, providing both with necessary state data. Superagent ContingencyAgenda Agenda Controller (list of suspended contingencies to i~e processed) (contains knowledge of muter production Proposed Schedule Agenda interactions) (ilsl of suspended proposed schedul~s) ¯ 4 ~ Cue8 nro~r~ edule Posting of proposed schedules with contingencies I Approval or Disapproval of proposed schedule I Response to negotiation Contingencies proposed to the individual agents w System Input Database (Cues Floor Sensors) Request for detailed data Detailed floor data Receiving Agent Shipping Agent Maintenance Agent Line 1 Agent Materials Handling Agent ¯ ¯ ¯ Line n Agent ProductionSystemAgents ~ ~(outpuO Schedule & Strategy changes Factory floor (inpu0 Superagent are entered by a sorting algorithmand pulled fxom the top of the list one at a time for processing.A The Superagent "watches" the system and has a contingency is the result of a newly-proposed controllingrole in that it triggers reschedulingof scheduled event; it is a changewhichmust take the subsystemsand has detailed knowledgeof the place in a related subsystem(e.g., release of master productionschedule. It embodiesknowledge committedresource) to accept the proposed new of the global goals of the productionsystem, the schedule. A proposedschedule mayhave a number local goals of the subsystems and how the of contingencies associated with it. All subsystemsinterconnect. This knowledge is used to contingenciesrelated to a proposedschedulemust decidewhichlocal goals are moreimportantat any be approved by the affected agent(s) or via given time to preserve the master production negotiation before the proposedschedule can be scheduleif a reschedule createsa local goalconflict. implemented. To manage rescheduling, the Superagent Negotiation is performedwhena contingency monitorscue data received fromthe shopfloor to creates a conflict, whichmaycause a proposed determineif any of the cue values are critical, scheduleto be rejected. Robustrepresentation and indicatinga possibleneedto rescheduleoneor more processingof the contingenciesis vital to improve subsystems. When sucha critical level is identified, the reliability of the rescheduled system a messageis sent to the affected subsystemagents. (Drummond, et al 1993). If a contingencycannot This messageis a request to reviewthe needfor a be approvedby the affected agent(s), the Superagent strategy and/or schedulechangeto return the cue has to use its knowledgebase to resolve the parameterto an acceptable level. Therequest may scheduling conflict. To preserve the master cause a revised schedule to be proposedby one or productiongoals either one of the subsystemswill more subsystems. If a proposed schedule is in have to yield its goals or both will have to conflict with the local goals of related subsystem compromise and change to an intermediate agents, negotiation is performedwith the help of schedule.In somecases, neither a single agent or a two agendas: one for keepingschedules proposed multiple agent compromisecan be reached, in by the subsystems and one for processing the whichcase the master productionschedulemust be contingenciesthat these proposedschedulescreate. compromised. Anagendain this context is a queuewhereitems . 35 3. Committee of Subagents Each agent contains a variety of scheduling strategies and a scheduling mechanism. The schedulingstrategy determinesthe techniqueusedto schedule subsystemtasks (may be analytical or heuristic), while the schedulingmechanism builds time-basedevent scheduleaccordingthe strategy. Theagent strives to maintainits local goals when altering the schedulealoneor both the scheduleand the strategy. AGENT STRUCTURE SUPERAGENT BOUNDING ASSUMPTIONS Schedule Clu~e 7 strategy AND/OR Change ? 4 SCHEDULING MECHANISM ~ Dc t tiled Shop Floor Information SYSTEM ] DATABASE H SCHEDUHNG STRATEGY A B C Propmed Schedule with Contingencies[ Oursystem contains agents whichrepresent the Automated Guided Vehicle (AGV) system, production lines, receiving, shipping, and maintenance. For agents that contain multiple schedulingstrategies to choosefrom,a requestfor a schedulechangemayinvolvea different strategy, a different scheduleor a combinationof both, as in the accompanyingfigure depicting a subsystem agent. Eachagent maintainsa time-basedschedule for its manufacturing subsystem. The Superagent sends the subsystem agent a request for reschedule based on cue parameters, along with the cue values. Theagent then solicits relevant detailed informationfrom the shopfloor databaseto determinewhethera schedulechangeis warranted. The agent determinesif a scheduling change,a strategy change,or both needto be made. The choice of scheduling strategy is based on a comparison of state conditionswith eachstrategy’s bounding assumptions, and on knowledgeof the conditionsunderwhicheach strategy performsbesL Theappropriate schedulingstrategy is determined and the scheduling mechanism develops the scheduleaccordingly.Relatedcontingenciesarise in two ways: the agent’s knowledgeof howcertain 36 changes in its subsystem propagate to related subsysi’dmsand notification of an inability to achieve prior commitmentsto other subsystems because of the proposedchange(s). The proposed schedule, along with any contingencies,is sent to the Superagent,whichwill then acceptor reject the schedule in the mannerdescribed in the previous section. If the schedule is accepted, it will be carried out as planned, otherwise the agent will revise its schedule and resubmit it to the Superagent. Ourpast research has focused on the material handling (AGV) subsystem. The remainingagents are currently underdevelopment. The AGV agent’s structure is comprisedof two sections: the AGV strategy selector and the AGV scheduler. The strategy selector is designed to dynamically adjust the AGVdispatching and mutingstrategy in real time suchthat eachstrategy is used whenit is most effective, resulting in improvedpart throughputand reducedAGV arrival time. The AGVagent’s rule-based reasoning utilizes output data from a series of empirical simulationexperiments(Interrante et al, 1993b). new strategy will be considered whenever a significant changein a parameterassociated with AGV performanceis detected (a cue parameter), such as a consistently low utilization value. The AGV scheduler handles the rescheduling of events becauseof contingencies imposedby other agents andminorinefficiencies, such as a productionline not receivingorderedparts. Thestrategy selector is a five-level structure whichis constructedas a decisiontree. Theswategy selector is triggered wheneverthe Superagent informs the AGVagent that there is a problem with a "cue" parameter.Thecue parametersinclude AGV utilization, AGV arrival time, localized line prioritization, localized order amount(to create congestion), level of backorders, AGV breakdown rate, production line breakdown rate, and overall order volume. The AGVagent then retrieves detailed informationon all of the cue parameters fromthe systemdatabaseanduses it as input to the decision tree. The parameters are analyzed and propagateddownthe tree basedupona set of rules and information drawn from the simulation experiments. Oncepropagationis complete, a new strategy is selected. This newly-cbosenstrategy will perform best under the current shop floor conditions. The AGVscheduler is triggered wheneverthe AGV agent receives a request for reschedule or a contingencyimposedby other agents. The agent accesses its database and checksthe AGV schedule and the availability of the AGVs.The AGV agent’s scheduleis reactionary andbased uponthe orders currentlyplacedfor part delivery.It is only madefor a short period of time. The schedule B. Reasoning Processes of cuesreachescritical levels, the Super’agent sends out re~fiests for rescheduling to the affected subsystems.Thecritical level for each cue depends on whether it occurs in combinationwith other critical cue values and is based on knowledge acquiredfromexperiencedmanufacturers(Interrante et al, 1993a).Therequest for reschedulingtells subsystemagentto evaluate its current performance and determinewhetheritsschedulingstrategy, its schedule,or both needsto be adjusted. 1. Active Rescheduling 2. Negotiation & Conflict Resolution contains knowledgeabout the assignmentsof each AGV. Withthis information, the agent determines its ability to carry out the request. It thensendits response to the Superagent,whichwill be one of the following:"yes, I can makethe change,""no, I cannot makethe changedueto lack of a particular resourceor a scheduleconflict," or "yes, I canmake the changeff a particular contingency is met." Thenew, as yet unaccepted,scheduledevelopedby Theterm "active rescheduling"is motivatedby the workof Swain,Bailard, and others (Swain,1991; a subsystemagent is sent to the scheduleagendaof BaUard,1991)in active vision, whereby"cues" are the Superagent. Theschedule is accompanied by a identified and used to enhanceperformancein the set of the previously-scheduledevents that were movedin rescheduling (which mayaffect other recognition of objects from visual images. Reschedulingis a non-value-addingactivity and agents). Suchchangesmaybe the result of the need shouldbe minimized.Yetthe reschcdulingfunction for a resource for an alternate purposethat its original intent. Theseevents are the contingencies is essential because of the dynamics of manufacturingsystems. Weemploymanufacturing that are mutedto the contingency agenda. Each contingencyis stampeduponreceipt to identify the parameters as cues to trigger a reschedule for time of receipt and the schedulewith whichit is particular subsystems.This approachsimplifies the process of deciding whichsubsystemneeds to associated. Theagendaprocessingmechanism sorts contingenciesby their id stamp,temporallocation be rescheduled based on current shop floor conditions. Onceit is determinedthat a reschedule and priority of the contingency.Contingenciesare is required,the actual reschedulingprocessinvolves removedfromthe stack one at the time, starting with the top-most contingencyat the time of the manymoremanufacturingparameters. Tochoosethe right time for changinga schedule, next processingcycle (i.e. everytime a contingency is finished processing,the current top-mostitem is the Superagent continuously monitors the cue parameters(Interrante et al., 1993a),an indication removedfrom the agenda)and processed. Theflow of which subsystems are in danger of running of information in the agenda system of the is depicted below: behindtheir schedule. Whenone or a combination Superagent Schedule fromLine 1 to AGVagent 1 scheduleproposal AGV contingency from from from Line 8 1 contingency AGV AGV contingency Line 8 0 0 0 0 0 0 to Line 6 agent _---to Maintenance agent agerlL toReceiving ~ AGVagent to Line2 8 contingency contingency 8 contingency to Receivingagent,~ to ShippingAgent lab.._ O The processing of a contingencyis performedby proposingthe changeto the affected sub-systems.If the change is not acceptable to one or more subsystem agents, the Superagent proceeds by beginning negotiation. It employsknowledgeof 37 the goals of each sub-system and the cost of proposedchangesin light of the affected goals. Usingthis informationthe conflict is resolved by either forcing oneof the systemsto accept the new eventandaccordinglyadjust the rest of its schedule (forfeiting its goal for the goodof the moreglobal "Intelligent Control of AGVs:Dynamic goal), forcing all involved systemsto reschedule Sc’fi-eduling via Selective Attention". the event in compromise,or relaxing the master Proceedings of the 1993 NSF Design and production schedule. Onceall contingenciesof a Manufacturing Systems Conference, Volume particular schedulehavebeenresolved,the schedule 2, 1993. agendais given permissionto resumethe suspended Intermnte, Leslie D., Daniel M. Rochowiak,and scheduleand return it to the senderagent as the Ben Sumeral. "Scheduling of Automated approvedschedule. Guided Vehicles for Material Handling: DynamicShop Floor Feedback". To appear in III. Conclusion ComputerControl of ManufacturingSystems. 1993. Muchresearch remains, including the development Law, Averil M. and W.David Kelton. Simulation of additional subsystemagents, the controlling of Modeling & Analysis. NewYork, McGrawpropagationof contingenciesand schedulechanges Hill, 1991. and the handling of contingency and schedule McKay,KennethN., "The Factory from Hell - A satisfaction in time. Boththe scheduleagendaand ModelingBenchmark",Intelligent, Dynamic the contingency agenda must be dynamically Scheduling for Manufacturing Systems updatedto reflect the fact that the processingof an Workshop,1993. agendaitem will alter the validity of someof the Nakasuka, Shiniehi and Taketoshi Yoshida, remainingitems. "Dynamic scheduling system utilizing machine Wehave described an architecture which is learning as a knowledgeacquisition tool", intended to model subsystem interactions in International Journalof ProductionResearch, manufacturingsuch that shop floor contingencies Vol. 30, No. 2, 1992. can be appropriately handled by an intelligent Romero,Iv(m S. and Nilda R. Romero."Chrysler schedulingsystem.Ourapproachaddressesthe need Acustar Simulation Model Development" to balancethe a priori bureaucraticstructure andthe Technical report, University of Alabamain a posteriori need to adapt to newand unforeseen Huntsville, Intelligent SystemsLaboratory, circumstancesin manufacturing. 1993. Smith, Stephen F., MarkS. Fox and Peng Si Ow, References "Constructing and Maintaining Detailed AutoSimulations, Inc. AutoModUser’s Manual ProductionPlans", A/Magazine,Fall 1986. and Lesson’s Manual. Star, S. L. "The structure of ill-structured Ballard, D. "Animate Vision," Artificial solutions: boundaryobjects and heterogeneous Intelligence, Vol48, 57-86, 1991. distributed problemsolving." In Gasser,L and BondandGasser,DistributedArtificial Intelligence. Huhns, M. (eds) Distributed artificial MorganKaufman,1988. intelligence Vol 2, London:Pitman, 37-53, Burke, P. and P. Prosser, "A distributed 1989. asynchronous systemfor predictive andreactive Swain, M. Proceedingsof the NSFActive Vision scheduling", Artificial Intelligence in Workshop,Chicago, IL, August, 1991. Engineering,Vol. 6, No.3, 1991. Wemer,E. "Cooperatingagents: A unified theory Drummond, Mark, John Bresina, Keith Swanson, of communication and social structure." "Contingent Scheduling: Opportunities and Research notes in Artificial Intelligence, Issues", Intelligent, Dynamic Schedulingfor Distributed Artificial Intelligence, Morgan ManufacturingSystems Workshop,1993. Kaufman,1989. Durfee, E., Lesser, V. and Corkill, D. "Trendsin Wilson, Carlton M., ThomJ. Hodgson, and cooperativedistributed problemsolving," IEEE Russell E. King, "A Note on Incorporation of Transactions on Knowledge and Data Improved Information in the Control of EngineeringI, 63-83, 1989. Multiple Vehicle AGVSystems". NCSU Egbelu, Pius J., and Jose" M. A. Tanchoco. TechnicalReport#90-14,North CarolinaState "Characterizationof AutomaticGuidedVehicle University, 1991. Dispatching Rules". International Journal Zlotkin, Gilad and Jeffrey S. Rosenschein, Production Research, Volume22, Number3, "Cooperation and Conflict Resolution via 359-374,1984. Negotiation AmongAutonomousAgents in Hewitt, C. "Offices are open systems." A CM NoncooperativeDomains",IEEETransactions Transactionson Office Information Systems, on Systems, Man,and Cybernetics, Vol. 21, Vol 4, No3, 271-287,1986. No. 6, November/December 1991. Interrante, Leslie D. Daniel M. Rochowiak,John S. Rogers, and Lysle E. Shields III. 38